QuantSR: Accurate Low-bit Quantization for Efficient Image Super-Resolution

NeurIPS 2023 · Spotlight

Haotong Qin, Yulun Zhang, Yifu Ding, Yifan Liu, Xianglong Liu, Martin Danelljan, Fisher Yu

Published paper | OpenReview | Citation

QuantSR performs accurate 2–4-bit image super-resolution through a Redistribution-driven Learnable Quantizer (RLQ) and a Depth-dynamic Quantized Architecture (DQA). It combines information retention in quantized operators with shared-weight depth choices for an accuracy/efficiency trade-off.

Published results

Selected ×4 results from Table 2. PSNR (dB) is higher-is-better; W/A denotes weight/activation bits in the quantized body. QuantSR-C uses SRResNet; QuantSR-T uses SwinIR-S. Training and evaluation follow Sections 3.4 and 4.1.

Backbone / method	W/A	Set5 PSNR	Urban100 PSNR
SRResNet / full precision	32/32	32.16	26.11
SRResNet / PAMS	4/4	31.59	25.32
SRResNet / QuantSR-C	4/4	32.00	25.88
SRResNet / PAMS	2/2	29.20	23.58
SRResNet / QuantSR-C	2/2	31.30	25.13
SwinIR-S / full precision	32/32	32.44	26.47
SwinIR-S / QuantSR-T	4/4	32.18	26.11
SwinIR-S / QuantSR-T	2/2	31.53	25.26

These are published reconstruction scores, not new runs. The paper's Params/Ops reductions (Table 3) use bit-width-aware theoretical accounting and depth choices; they do not establish equal reductions in GPU latency or runtime memory. Head/tail layers are retained at full precision in the reference protocol.

What this paper supports

• RLQ mitigates quantization-induced representation homogeneity using redistribution in forward and backward propagation (Section 3.2; Table 1).
• DQA supports inference at multiple depths with shared weights (Section 3.3; Figure 4).
• For SRResNet at ×4 and W2A2, QuantSR-C improves Set5 PSNR by 2.10 dB over PAMS (Table 2).
• The approach works on the evaluated convolutional and Transformer SR backbones, with remaining full-precision gaps shown above (Section 4.3; Table 2).
• Bit-width and depth jointly control theoretical storage/computation and reconstruction quality (Table 3).

Dependencies

# Go to the default directory
pip install -r requirements.txt
python setup.py develop

Execution

# We provide script to test our 4-bit QuantSR-C
sh test.sh

Citation

Please cite the published paper below. Open paper versions are linked at the top of this README.

@inproceedings{qin2023quantsr,
  title = {{QuantSR}: Accurate Low-bit Quantization for Efficient Image Super-Resolution},
  author = {Haotong Qin and Yulun Zhang and Yifu Ding and Yifan Liu and Xianglong Liu and Martin Danelljan and Fisher Yu},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2023},
  volume = {36},
  url = {https://proceedings.neurips.cc/paper_files/paper/2023/hash/b2169d573d75ff90c7b12dc3a5fc2898-Abstract.html}
}